An high order relaxation labeling neural network for feature matching

The main aim of this work is to propose a new artificial neural network architecture to solve the motion correspondence problem by a nonlinear relaxation labeling approach estimating correct matches from high order compatibility measurements. High directional variance features extracted from a frame at time t and the corresponding high correlated ones of the frame at time t + 1 are used to determine respectively two relational graphs with five-order links labeled by the geometrical invariant value of cross-ratio of any Ave coplanar features (relational graph nodes). Cross-ratio similarities between relational graphs are used as constraints to determine compatibilities between feature matches (association graph nodes). We map the recovered association graph into an appropriate neural network architecture. The feature matching problem is then solved by relaxating the neural network according to dynamical equations following an heuristic nonlinear relaxation scheme. Being based on geometrical invariance of coplanar points as main constraint, the approach recovers matches only for a set of coplanar features. Actually, this is not a loss of generality, because in a lot of indoor real contexts the viewed scene can be well approximated to a plane. Finally, In our experimental tests, we found our method to be very fast to converge to a solution, showing as higher order interactions help to speed-up the process.